Handle empty ECS Clusters properly

[jeffreykoetsier]

This PR fixes a bug in the ECS provider.

When the ECS Provider is configured to poll multiple clusters, it'll show unexpected behavior when one of the clusters has 'zero' tasks running.

1. If the first cluster Traefik encounters has no running tasks the listInstances function is ended immediately by returning an empty slice of ecsInstances. Successive clusters are simply ignored.

2. If the 2nd or up cluster has no tasks running the listInstances function will continue trying to fetch data for these non-existing-tasks. The 'early return' condition will not detect this due to counting the length of taskArns slice. This slice var is used by each cluster discovery without being emptied, so it'll be filled with taskArns from the previous cluster. This eventually results in incomplete API calls to AWS ECS.

In the 2nd case Traefik (lookupEc2Instances) sends a 'DescribeContainerInstances' request to AWS with an 'empty' list of containerArns. AWS returns a 4xx.

time="2017-09-19T14:27:49Z" level=error msg="Provider connection error InvalidParameterException: Container instance cannot be empty.
        status code: 400, request id: b6c38d74-9d46-11e7-9fc8-3d440bc8c23f, retrying in 388.555963ms"

Fix:

• Move initialization of taskArns slice into the for loop. Each iteration of the loop now starts with an empty slice.
• Do not return from the function when len(taskArns) == 0. Just continue to the next iteration of the loop.

Fixes #2168

[mmatur]

LGTM 👍

[ldez]

LGTM

[juliens]

LGTM

[emilevauge]

LGTM